California AI Safety|EU Regulation: LLMs Emergent Abilities and Existential Threat

Questions that may be essential to AI regulation for now may include current and potential misuses, sources of those misuses, and why they are technically permissible.

Broad AI regulations like those for energy, pharmaceuticals, and other sectors may apply in some form, but ultimately, AI regulation that may work would likely be technical. Physical laws can regulate digital products but would find limitations.

There are several AI tools that misuses are possible from that may slip regulations. There are ways that some frontier AI models are self-regulating that may leave little room for adjustments from authorities. Even in the event of presenting models for safety evaluation before public release, frontier AI companies may have done some work, covering what they expect would be sought.

How can AI regulation technically fight misuse? This includes the possibility to explore real-time multimodal misuses, across AI models. How is it possible to track and monitor some outputs of all AI tools that are available in a jurisdiction?

California Senate has SB 1047 to regulate AI about "Giving companies more flexibility in how they meet their responsible development obligations while still holding them liable if irresponsible behavior leads to catastrophic harm. Allowing a state agency to change the compute threshold for a model to be covered under the bill starting in 2027, rather than keeping it fixed in statute at 10^26 flops. The requirement that the model cost at least $100 million to train cannot be changed. Requiring companies get third-party safety audits by 2028. Strengthen whistleblower protections. Requiring developers of such large “frontier” AI models to take basic precautions, such as pre-deployment safety testing, red-teaming, cybersecurity, safeguards to prevent the misuse of dangerous capabilities, and post-deployment monitoring. "

The European Commission in AI Act enters into force, states that "On 1 August 2024, the European Artificial Intelligence Act (AI Act) enters into force. The Act aims to foster responsible artificial intelligence development and deployment in the EU. Proposed by the Commission in April 2021 and agreed by the European Parliament and the Council in December 2023, the AI Act addresses potential risks to citizens’ health, safety, and fundamental rights. It provides developers and deployers with clear requirements and obligations regarding specific uses of AI while reducing administrative and financial burdens for businesses. The AI Act introduces a uniform framework across all EU countries, based on a forward-looking definition of AI and a risk-based approach:"

The problem with AI regulation is not that laws are not possible around it or that industry standards cannot be set. It is just that the ability to catch misuses on the go is better than regulation at the gate.

Simply, it is vital to catch misuses, which in this case goes beyond the aerial view of frontier companies or start-ups but to tools, across usage sources, consistently. This will be like fighting digital piracy at any source it is available within a jurisdiction, not just on major platforms, or pointing to physical laws when digital evasion is knavish.

There are research options like penalty-tuning and web scraping of available AI tools. Penalty-tuning could be a way to penalize LLMs, by compute, usage, or language, in a way that the model can know, if it outputs misinformation or deepfakes.

Web scraping could be possible for AI tools that are available in search results so that some of the keywords around misuses can be tracked to expect them at their likely destinations, to prevent the harm they might cause.

AI regulation is a problem of research first, then of laws. Exploring how to advance, technically, against misuses could guide what directions laws may take.

There is a recent preprint on arXiv, Are Emergent Abilities in Large Language Models just In-Context Learning?, where the authors wrote, "We present a novel theory that explains emergent abilities, taking into account their potential confounding factors, and rigorously substantiate this theory through over 1000 experiments. Our findings suggest that purported emergent abilities are not truly emergent, but result from a combination of in-context learning, model memory, and linguistic knowledge.

Our work is a foundational step in explaining language model performance, providing a template for their efficient use and clarifying the paradox of their ability to excel in some instances while faltering in others. Thus, we demonstrate that their capabilities should not be overestimated."

AI is simply not an existential risk because of prediction-based LLMs. In the human mind, what is called prediction is a type of distribution or relay. LLMs have a substantial part of that, but they do not have what follows after [the incoming half of electrical signals, conceptually].

As AI gets better, with non-concept features and distributions similar to how relays within human memory result in human intelligence, they may pose higher levels of risks.